Thematic analysis with ChatGPT - 3 ways to create and/or organize your themes in ChatGPT

ChatGPT can speed up thematic analysis in three practical ways—first by drafting themes from a code list, second by sorting codes into already-chosen theme buckets, and third by clustering long lists of subthemes into cleaner categories. The most useful pattern across all three is keeping final interpretive control: ChatGPT can reorganize and propose structure, but the researcher still decides what themes mean and which ones best tell the study’s story.

The first approach—asking ChatGPT to generate themes directly from codes—is presented as the least preferred option. The core concern is that themes in qualitative research aren’t supposed to “emerge magically” from data; they’re selected and shaped by the researcher’s knowledge, study focus, and conceptual priorities. Still, the method can be demonstrated by providing ChatGPT with context: a hypothetical study of educational leaders’ lived experiences during crises (conflict, war, economic crisis, and other disruptions), plus a list of codes representing challenges, strategies, and suggestions. With that setup, ChatGPT returns candidate themes and assigns codes under each one. In the example, it produces broad, polished-sounding themes such as “Institutional preparedness and resilience,” “Equitable and Effective Education delivery,” and “Community collaboration,” along with code allocations. The critique is that these themes can come out too general or abstract, mixing different kinds of content (good practices, challenges, and recommendations) rather than separating them into more actionable groupings like challenges, strategies, and suggestions.

The second approach is more aligned with how many researchers work: start with predefined themes and use ChatGPT to group codes into those categories. Here, the prompt is narrower—asking ChatGPT to place the same code list into buckets such as “challenges,” “strategies to overcome challenges,” and “suggestions.” This tends to be more convenient than generating themes from scratch, but it may still be of limited value if the researcher already understands their codes well. The example also notes that ChatGPT can struggle when codes are less descriptive than the model expects, leading to imperfect categorization that still requires human review.

The third approach—grouping subthemes into higher-level categories—is described as the most consistently helpful. When a researcher has a long list of challenges (or strategies, or suggestions), ChatGPT can cluster them into several categories, producing a more readable structure. In the demonstration, a list of challenges gets grouped into categories such as “financial and resource challenges” and “stuffing and personal issues” (as generated in the example). The output can include overlaps or awkward category boundaries, but it often provides a useful starting point. Researchers can then refine the number of groups, adjust labels, and resolve overlaps, using ChatGPT to maximize the value of their own work rather than outsourcing interpretation.

Overall, the practical takeaway is staged assistance: use ChatGPT for organization and scaffolding—especially for clustering and reformatting—while keeping the thematic decisions, definitions, and narrative framing firmly in human hands.